Who actually needs this?

Organisations where somebody outside them requires it, which is the usual signal. Adopting it out of general caution takes on a permanent operational burden in exchange for a guarantee that a contractual commitment and a controlled deployment would have provided at a fraction of the cost.

What does it rule out?

Anything that reaches outward: looking something up, calling another service, fetching current information. That is a larger constraint for AI than for most software, so the applications that work are the ones answering entirely from material already inside the boundary.

What is the real cost?

People and procedure rather than equipment, which is the reverse of how it is usually budgeted. Somebody owns bringing things in, keeping the environment current enough to be safe, and being available when it breaks with no vendor able to look at it remotely. Equipment is the visible cost and staffing is the one that decides whether the arrangement is sustainable.

Glossary

Air gapped

A deployment with no connection to the outside world at all, used where material must not leave the premises and where updates have to be carried in.

In plain terms

The machines are not connected to anything outside. No updates arrive on their own, nothing can be looked up, and anything that needs to get in or out is carried by a person following a procedure. For AI that means a model and everything around it living entirely inside a boundary that nothing crosses electronically.

Why it matters

Because it is the strongest answer available to where does our material go, and because the cost is almost entirely in living with it rather than in setting it up. Organisations that need it usually know, because somebody outside them requires it. Organisations that adopt it out of caution take on a permanent and considerable burden in exchange for a guarantee a contract would have given them.

How it works

Everything that would ordinarily arrive automatically now arrives by hand. Model updates, security patches and any change to the software have to be brought in through a controlled process, which is deliberate and slow by design. The slowness is the point rather than a defect, and it is also the reason these deployments drift further behind than their operators intend.

The model is fixed until somebody changes it, which cuts both ways. Behaviour will not shift underneath you, which is a genuine advantage for anything that has been carefully validated. Improvements will also not arrive, so a deployment can sit for a long time on a model that has been superseded, and nobody outside notices.

Nothing that needs to reach outward will work at all, and that is more limiting for AI than it is for most other software. Anything that looks something up, calls another service or fetches current information is unavailable by construction, so the useful applications are the ones that work entirely from material already inside the boundary.

The running cost is people and procedure rather than equipment. Somebody has to own the transfer process, keep the environment current enough to be safe, and be available when it stops working, with no vendor able to look at it remotely. That is the line that decides whether an organisation can sustain the arrangement, and it is the one estimated most optimistically.

How far the boundary goes

The gap between the middle mark and the right-hand one is much larger than it appears, and it is the distinction this page exists to make. Both are your hardware in your building, both remove the questions about what a provider does with your material, and they look similar in a diagram. The difference is that the middle mark still participates in the ordinary flow of software: patches arrive, licences check in, a vendor can look at a problem with you. The right-hand mark opts out of all of that, permanently, and replaces each one with a procedure somebody owns. That is why the honest question when this is proposed is not whether the organisation can build it but whether it can still be running properly in three years, because the failure mode is not a breach. It is a slow drift into an environment nobody has updated and nobody quite wants to touch.

Seen in the wild

Running an open-weight model on a machine with no outward connection, which is the ordinary shape of this.
Ollama
A desktop application loading models from local files, so nothing needs to be fetched to use one.
LM Studio
A chat interface and document store running entirely against local material, so a disconnected team still has something usable.
AnythingLLM

Common misconceptions

People assume

It is on-premises with extra care.

In fact

It is a different operating model. On-premises hardware still reaches out for updates, licences and support; this reaches out for nothing, so every one of those becomes a manual procedure somebody owns. The setup resembles on-premises and living with it does not.

People assume

The hard part is building it.

In fact

The hard part is the years afterwards rather than the build itself. Keeping a disconnected environment current, staffed and functioning without any remote assistance is continuing work, and deployments that quietly drift out of date are the normal outcome rather than the exceptional one.

Telling them apart

Air gapped vs On premises

Air gapped

No connection at all. Everything arrives by hand.

On premises

Your hardware, still connected for updates, licences and support.

Ask how a security patch arrives. If the answer involves a person carrying it, this is the one.

Questions

Who actually needs this?: Organisations where somebody outside them requires it, which is the usual signal. Adopting it out of general caution takes on a permanent operational burden in exchange for a guarantee that a contractual commitment and a controlled deployment would have provided at a fraction of the cost.
What does it rule out?: Anything that reaches outward: looking something up, calling another service, fetching current information. That is a larger constraint for AI than for most software, so the applications that work are the ones answering entirely from material already inside the boundary.
What is the real cost?: People and procedure rather than equipment, which is the reverse of how it is usually budgeted. Somebody owns bringing things in, keeping the environment current enough to be safe, and being available when it breaks with no vendor able to look at it remotely. Equipment is the visible cost and staffing is the one that decides whether the arrangement is sustainable.

Key takeaways

Everything that would arrive automatically now arrives by hand, by design.
The model is fixed, so behaviour is stable and improvements do not come either.
Anything that reaches outward is unavailable, which limits AI more than most software.
The cost is people and procedure over years, not the build.